Image Colorization Based On Multi-Skip Connections And Color-UNET++

In our daily life,color not only makes our world colorful,but also expresses and presents many important information.Nowadays,coloring images by computer is becoming more and more popular.With the development of artificial intelligence,the use of deep learning for image coloring has become a development trend,but the existing deep learning methods are facing many problems,such as insufficient color saturation,low image clarity,and difficult training.The problem of how to achieve a fully automatic deep learning method that is easy to train and can be colored well has become an important problem in the field of computer vision.In this paper,two image colorizing methods based on deep learning are proposed,namely,the image colorizing algorithm based on multi-skip connection and the image colorizing algorithm based on Color-UNET++.The two methods are used to carry out image coloring contrast experiments on different data sets to test the effect and performance of the two algorithms.The main contributions of this paper are:1.By reference to the design idea of jump connection of residual network,an image colorizing method based on multi-skip connection model is proposed.Among them,(1)the model structure of multi-skip connection is designed to make it have stronger learning ability and feature extraction effect when learning image color information.(2)According to the horizontal and vertical corresponding pixel gradient losses of the image,the corresponding pixel gradient loss function was designed and combined with the mean square error loss function to form a compound loss function.Through comparative experiments,it was proved that the image color optimization and learning ability were greatly improved.2.An image colorizing method based on Color-UNET++model is proposed.Among them,(1)a transfer layer of intermediate skip connection is added to the corresponding structure of the down-sampling and the up-sampling of the deep convolutional network structure to solve the problem of information loss between the encoder and the decoder.(2)The scheme of adding a special convolutional block in the up-sampling stage of the network structure is designed,which successfully solves the problem of checkerboard artifacts caused by deconvolutional network.3.The two methods proposed in this paper both adopt the method of color space conversion,converting RGB color space into YUV color space,thus greatly improving the color effect and image clarity of the model.Compared with similar methods at home and abroad,it is proved that the algorithm proposed in this paper can effectively improve the quality of image colorizing and has a considerable degree of reliability through the comparison of coloring images and evaluation results of objective indicators.